Water Segmentation for Unmanned Ship Navigation Based on Multi-Scale Feature Fusion

The segmentation of the navigation area from water images is of great significance in the safe and automated navigation of unmanned vessels. However, accurate segmentation of water boundaries in real time under the interference of water-surface light changes is still a challenging problem. In this study, a river water segmentation method based on a multi-scale feature fusion network (Fuse-ResNet), which improves the segmentation of river water boundaries by focusing more on the target features, is proposed. An Atrous Spatial Pyramid Pooling (ASPP) mechanism is introduced to expand the sensory field and enhance the relevance of contextual information at different scales. Experimental results based on the USVInland unmanned vessel dataset show that the proposed method achieves a mean Pixel Accuracy (mPA) of 99.25% and a mean Intersection over Union (mIoU) of 98.45% in the water category, which are better than those achieved by mainstream models. The interference of background noise resulting from water surface reflection can be suppressed to improve the accuracy of water segmentation for unmanned vessel navigation.